Faculty-supervised AI-supported BOPPPS-CBL and performance on a simulated neurological-emergency assessment among junior-college medical interns: a batch-stratified randomized educational study

County hospitals and township health centers in China provide frontline care for time-sensitive neurological emergencies, and three-year junior-college clinical medicine programs are an important pathway into this workforce. We evaluated a faculty-supervised artificial intelligence (AI)-supported BOPPPS (Bridge-in, Objectives, Pre-assessment, Participatory Learning, Post-assessment, and Summary) and case-based learning (CBL) instructional package among junior-college medical interns.

In this prospective, batch-stratified randomized educational study, 118 interns from 12 monthly rotation batches were assigned to AI-supported BOPPPS-CBL (n = 58) or conventional BOPPPS-CBL (n = 60) during a one-month neurology rotation. The intervention used standardized DeepSeek prompts for Socratic questioning, error-focused feedback, and structured reflection under faculty supervision. The primary endpoint was the total score on a locally developed four-station simulated neurological-emergency assessment (0-100). Secondary outcomes included a written examination, while learner-reported questionnaire items were exploratory. Regression models adjusted for rotation batch, baseline academic performance, and digital literacy. The study was not prospectively registered; no public trial registration identifier or registry URL exists, and no publicly time-stamped protocol or statistical analysis plan was available.

All 118 participants completed the primary assessment. Mean total scores were 82.07 in the AI-supported group and 76.43 in the control group, with an adjusted between-group difference of 5.67 points (95% CI, 2.86-8.49; P < 0.001; adjusted Cohen's d = 0.74). In secondary station-specific analyses, only lesion localization met the multiplicity-adjusted threshold (adjusted mean difference, 2.58 points; 95% CI, 1.25-3.91; P < 0.001; adjusted d = 0.71). A Station 2-specific ICC was not estimated. Post hoc sensitivity analyses excluding the station containing reused case contexts yielded standardized estimates of similar magnitude but could not exclude cueing or carry-over. Written-examination estimates were smaller and imprecise. Exploratory learner-reported findings indicated greater perceived preparation-time burden in the AI-supported group.

Assignment to the faculty-supervised AI-supported BOPPPS-CBL package was associated with higher scores on a locally developed simulated neurological-emergency assessment than conventional BOPPPS-CBL. These findings do not establish transfer to workplace clinical performance or patient outcomes. Multicenter studies using independent assessment cases, longer-term follow-up, and workplace-based outcomes are warranted.